In a new study, Yoan Hermstrüwer and David Imhof analyze how AI can help antitrust authorities predict cartels by evaluating international bidding data from countries with similar legal and market structures.
Cartel busted in Switzerland. The other was dismantled in Finland. And more bidders were indicted in Brazil and the United States. Each incident generates valuable data about how colluding companies behave: statistical fingerprints of obvious patterns and coordination in bidding. But will the lessons learned in Zurich help catch the riggers in Los Angeles?
The implications are particularly significant for government budgets. Public procurement accounts for approximately 13% of GDP in member countries of the Organization for Economic Co-operation and Development, and approximately 30% in developing countries, and the damage caused by bid rigging is staggering. Cartels typically jack up government contract prices by anywhere from 20 to 50 percent, draining taxpayer money that would have gone toward roads, schools, and hospitals. Enforcement agencies are increasingly turning to machine learning, and more generally artificial intelligence, to assess patterns in data and counter collusion. A growing number of antitrust authorities around the world are now considering algorithmic screening tools.
The promise is appealing. If AI models trained on known cartel behavior in one jurisdiction or industry can reliably warn of collusion in another jurisdiction or industry, resource-strapped government agencies could pool data and increase detection power. But the risks are just as obvious. Models that cannot reliably generalize data can flood investigators with false alarms or miss real cartels hiding behind unfamiliar legal and institutional rules.
In our new paper, “Predicting cross-jurisdictional machinery in antitrust,” we provide systematic evidence on how cartel detection models can be used to predict collusion across borders and different industries. We refer to the quality of a model that can be accurately applied to different jurisdictions and industries as “transferability.” Our findings reveal both the promise and limitations of such models.
First, cross-jurisdictional cartel detection is possible, with accuracy greater than 85% in institutionally similar markets. Second, the new bidder-level screen, which tracks companies across multiple auctions, significantly outperforms traditional bidder-level approaches that only treat offers individually. Third, the transferability of machine learning models is undermined when procurement rules, contract types, and market structures differ significantly between jurisdictions. Specifically, models applied to a variety of industries (e.g., from construction to the milk market) fail almost completely and perform worse than random guesses.
How to measure transferability
Detecting collusion is fundamentally a classification problem. Given the statistical properties of bids, the challenge is to predict whether bids are colluding or competing. We employ ensemble machine learning techniques that combine multiple algorithms to improve prediction accuracy and robustness. Specifically, we train three different models: Random Forest, Gradient Boosting, and XGBoost, and use different techniques to aggregate their predictions. By combining these models, you can leverage their complementary strengths while mitigating their individual weaknesses.
We compile bidding data from documented cartel cases across six jurisdictions: Switzerland, Finland, Sweden, Japan (Okinawa), Brazil, and the United States (California and Ohio). This dataset contains approximately 30,000 individual bids from public procurement auctions, primarily in road construction and civil engineering.
Our approach combines traditional screening, recommended in the latest version of the OECD Guidelines on Anti-Collusion in Public Procurement, in the form of statistical indicators that flag suspicious bidding patterns (such as bid dispersion in tenders), with new screening methods in the form of bidder-level measurements that track how individual companies behave over time in multiple procurement auctions. While existing methods treat each bid individually, our bidder-level screening captures persistent coordination patterns that emerge when the same firms repeatedly collude.
We test four model specifications, each capturing how firms bid over different time periods. The first uses only bid-level screens. That is, measures such as bid dispersion, skewness, and the spread between the highest and lowest bids within a single procurement auction. The second incorporates a moving average to track each company’s bidding pattern over the past five bids. The third uses bidder-level averages over time to calculate each firm’s typical behavior across all observed auctions. The fourth is a combination of medium-term and long-term measures. This design allows us to assess whether persistent behavior patterns are more informative than snapshot observations.
To test cross-jurisdictional transferability, we use a design that excludes one jurisdiction. That is, we train the model on data from all but one jurisdiction to predict collusion in the retained jurisdiction. This approach simulates what would happen if an enforcement agency attempted to use foreign cartel data to screen its home market.
A key challenge in applying machine learning to antitrust law is the black box problem. Complex models may achieve high accuracy, but may not reveal the reason for flagging a particular bid as suspicious. We address this using SHAP (SHApley Additive exPlanations). This is a cooperative game theory-based method that decomposes each prediction into its individual feature contributions. SHAP values reveal which statistical screens are driving the model’s decisions, both globally (the most important feature in all predictions) and locally (why a particular bid was flagged). This transparency is essential in antitrust law, and government agencies must be able to explain and defend their review decisions.
Main findings
Cross-jurisdictional detection works very well when procurement rules and contract types match. In the case of Finland, our model correctly classifies almost 90% of bids as bid-rigging or competition, despite being trained entirely on foreign data. Sweden and Switzerland have similarly strong results, with accuracy exceeding 75%. These jurisdictions share similar bidding procedures, contract types, and market structures, primarily related to road construction.
The long-term benefits from tracking companies are significant. In Finland, accuracy jumps from 76 percent using bid-level reviews alone to nearly 90 percent when bid-level measurements are used. This improvement reflects the fundamental insight that collusion is not a single bidding incident. It is a dynamic process rooted in repeated interactions. Companies that coordinate do so tenaciously, and our screens capture this pattern.
Japan provides a striking example of how specific procurement methods and auction designs can impact model performance. Japanese bidding rules impose explicit upper and lower limits on acceptable bids, effectively compressing the range of observable prices. This truncation mechanically reduces the bid variance. This is the very signal our screens use to distinguish between collusion and competition. As a result, competitive bidding in Japan is statistically similar to collusive bidding in other countries, resulting in a large number of false positives. Therefore, training a model based on Japanese data does not help accurately predict collusion in Finland, Sweden, or Switzerland.
The most notable limitations appear when the model crosses sector boundaries. Our Brazilian cartel data comes from oil infrastructure contracts, not road construction. These projects are larger, more complex, and more heterogeneous, which naturally leads to greater dispersion in bids, even under competition. Models trained on construction data fail to systematically detect bid rigging in Brazil and incorrectly interpret wide variations in bids as normal competitive variations. The lesson is straightforward. Products are important in machine learning models. Paving is not oil. It would be misleading to pretend otherwise.
The Ohio school milk cartel of the 1980s illustrates an even more fundamental barrier. This collusive arrangement relied primarily on bid suppression and market allocation, a strategy that reduces the number of bids rather than distorting the allocation of bids. With fewer bidders per bid and standardized products, you won’t see any of the statistical fingerprints found in construction sector inspections. The model’s performance is worse than a coin toss.
Policy lessons
Cross-border collaboration is possible, but there are caveats. Enforcement agencies can make meaningful use of cartel data from other jurisdictions, but only if the markets share a similar institutional framework. For the European road construction market, our results suggest that pooling data across jurisdictions could significantly improve detection power. The OECD’s commitment to international cooperation in cartel examinations finds empirical support here.
Organizational awareness is essential. Sourcing rules are not neutral; they shape the very data that discovery models analyze. Features such as bid price caps, eligibility requirements, and contract bundling practices can systematically alter statistical patterns in ways that confuse algorithmic screening. Authorities deploying AI tools need to understand how local rules interact with model assumptions.
The failure of cross-sectoral transfers is perhaps our most alarming finding. A model that works almost perfectly on a construction site may be useless, or worse, misleading, on a medical procurement or food supply contract. Different industries have different cost structures, competitive dynamics, and collusion mechanisms. There is no one-size-fits-all solution.
The model is a risk indicator, not a judgment. Even our best specifications can produce false positives and miss true cartels. These tools should flag suspicious patterns for human review rather than triggering automatic enforcement actions. The purpose is to help authorities allocate scarce investigative resources more efficiently, not to replace prosecutors or human judgment more generally.
The scope of algorithmic cartel testing will only expand as procurement systems become digital and generate richer audit trails. Our study suggests that this is a promising frontier, but the institutional context requires close attention. Machine learning can help catch cross-border cartels, but only if you understand the legal rules and practices that apply in each market.
Author Disclosure: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of the authors and do not necessarily represent the opinions of the University of Chicago, the Booth School of Business, or their faculty.
Stay up to date on ProMarket’s coverage of politics and economics and other content from the Stigler Center by subscribing here to ProMarket’s weekly newsletter, Special Interest.
